Before the successful renaming, a session accessed the ghost table, which had already unlocked the original table.

There is a very small probability that other sessions dml operations on the original table will occur,
and this dml operation will appear in the original table after renaming, resulting in data loss.

Author: dbking

go/logic/applier.go

@@ -1099,7 +1099,7 @@ func (this *Applier) RevertAtomicCutOverWaitTimeout() {
 }
 
 // AtomicCutOverMagicLock
-func (this *Applier) AtomicCutOverMagicLock(sessionIdChan chan int64, tableLocked chan<- error, okToUnlockTable <-chan bool, tableUnlocked chan<- error) error {
+func (this *Applier) AtomicCutOverMagicLock(sessionIdChan chan int64, tableLocked chan<- error, okToUnlockTable <-chan bool, tableUnlocked chan<- error, renameLockSessionId *int64) error {
 	tx, err := this.db.Begin()
 	if err != nil {
 		tableLocked <- err
@@ -1190,6 +1190,23 @@ func (this *Applier) AtomicCutOverMagicLock(sessionIdChan chan int64, tableLocke
 		// We DO NOT return here because we must `UNLOCK TABLES`!
 	}
 
+	this.migrationContext.Log.Infof("Session renameLockSessionId is %+v", *renameLockSessionId)
+	// checking the lock holded by rename session
+	if *renameLockSessionId > 0 {
+		for i := 0; i <= 50; i++ {
+			err := this.ExpectMetadataLock(*renameLockSessionId)
+			if err == nil {
+				this.migrationContext.Log.Infof("Rename session is pending lock on the origin table !")
+				break
+			} else {
+				if i == 50 {
+					return err
+				}
+				time.Sleep(10 * time.Millisecond)
+			}
+		}
+	}
+
 	// Tables still locked
 	this.migrationContext.Log.Infof("Releasing lock from %s.%s, %s.%s",
 		sql.EscapeName(this.migrationContext.DatabaseName),
@@ -1409,3 +1426,27 @@ func (this *Applier) Teardown() {
 	this.singletonDB.Close()
 	atomic.StoreInt64(&this.finishedMigrating, 1)
 }
+
+func (this *Applier) ExpectMetadataLock(sessionId int64) error {
+	found := false
+	query := `
+		select /* gh-ost */ m.owner_thread_id
+			from performance_schema.metadata_locks m join performance_schema.threads t 
+			on m.owner_thread_id=t.thread_id
+			where m.object_type = 'TABLE' and m.object_schema = ? and m.object_name = ? 
+			and m.lock_type = 'EXCLUSIVE' and m.lock_status = 'PENDING' 
+			and t.processlist_id = ?
+	`
+	err := sqlutils.QueryRowsMap(this.db, query, func(m sqlutils.RowMap) error {
+		found = true
+		return nil
+	}, this.migrationContext.DatabaseName, this.migrationContext.OriginalTableName, sessionId)
+	if err != nil {
+		return err
+	}
+	if !found {
+		err = fmt.Errorf("cannot find PENDING metadata lock on original table: `%s`.`%s`", this.migrationContext.DatabaseName, this.migrationContext.OriginalTableName)
+		return this.migrationContext.Log.Errore(err)
+	}
+	return nil
+}

go/logic/migrator.go

@@ -669,8 +669,9 @@ func (this *Migrator) atomicCutOver() (err error) {
 	lockOriginalSessionIdChan := make(chan int64, 2)
 	tableLocked := make(chan error, 2)
 	tableUnlocked := make(chan error, 2)
+        var renameLockSessionId int64
 	go func() {
-		if err := this.applier.AtomicCutOverMagicLock(lockOriginalSessionIdChan, tableLocked, okToUnlockTable, tableUnlocked); err != nil {
+		if err := this.applier.AtomicCutOverMagicLock(lockOriginalSessionIdChan, tableLocked, okToUnlockTable, tableUnlocked, &renameLockSessionId); err != nil {
 			this.migrationContext.Log.Errore(err)
 		}
 	}()
@@ -735,6 +736,7 @@ func (this *Migrator) atomicCutOver() (err error) {
 	// Now that we've found the RENAME blocking, AND the locking connection still alive,
 	// we know it is safe to proceed to release the lock
 
+        renameLockSessionId = renameSessionId
 	okToUnlockTable <- true
 	// BAM! magic table dropped, original table lock is released
 	// -> RENAME released -> queries on original are unblocked.